Agent 工程

GPT-6 增加 cache breakpoint、诊断、预热和可变配置。长跑 Agent 需要主动保持共享前缀稳定，并监控缓存命中率。

Claude Opus 5.5 agent API compatibility, RRSI controls for harness overfitting, Harness-Zero behavior distillation, and Strands harness defaults.

本期收录 Claude Opus 5.5 的 Agent API 兼容变化、RRSI 的 Harness 过拟合控制、Harness-Zero 的行为蒸馏，以及 Strands harness 的默认策略。

LangGraph human-approval schemas, Deep Agents safeguards for parallel writes to the same file, and tighter OAuth boundaries in the MCP Ruby SDK.

本期收录 LangGraph 人工审批结构约束、Deep Agents 同文件并行写保护，以及 MCP Ruby SDK OAuth 边界收紧。

This Minecraft agent uses Astra for slow planning, JEV for frequent action selection, and Mineflayer for execution, with logs and scripts to verify runs.

这个 Minecraft Agent 让 Astra 负责慢规划、JEV 负责高频动作选择、Mineflayer 负责确定性执行，并用日志和验证脚本保留运行证据。

How task dependencies shape multi-agent results, with SAIGE experiments, ablations, and OpenAI guidance for deciding when to delegate.

Rethinking Multi-Agent Collaboration 说明多 Agent 的收益取决于任务依赖结构。本文结合 SAIGE 主实验、消融实验和 OpenAI 多 Agent 工程建议，整理可执行的委派判断。

OpenAI Agents SDK runtime-boundary fixes, stage-aware memory retrieval in AdaRepair-Mem, and static configuration checks with Agent Lint.

本期收录 OpenAI Agents SDK 运行时边界修复、AdaRepair-Mem 阶段化记忆检索，以及 Agent Lint 静态配置检查。

TypeSafe Jev uses Choice, Score, and Noul for constrained agent decisions while code retains control of valid actions and side effects.